The Mystery Talk David E' Smith NASA Ames Research Center desmitharc'nasa'gov

1
The Mystery Talk David E. SmithNASA Ames
Research Centerdesmith_at_arc.nasa.gov
2
AIPS 2000
Coping with Time Continuous Quantities
3
Classical Planning Assumptions
Instantaneous actions No time constraints No
resources No continuous quantities No optimization
Bleak?
4
Temporal Planners
Zeno (Penberthy) Trains (Allen) Descartes
(Joslin) IxTeT (Ghallab) HSTS (Muscettola) Europa
(Jonsson et al.)
5
Critical Issues
6
AIPS 2002 Planning Competition
No SAT planners No ILP planners LPG (local
search on PG) FF, SAPA (forward SS
search) MIPS VHPOP (POCL)
7
Significant Progress
distance heuristics
8
The Top 6 Reasons NASA Cant Use those
Impressive New Planning Systems(LPG, FF, SAPA,
MIPS, VHPOP, TLPlan, TALPlanner)
9
Reason Number 1

• Exogenous conditions events
• Communication windows
• Observation windows
• Lighting conditions
• Orbit insertion
• Celestial events
• Time constraints EVERYWHERE !

10
Exogenous Events
Solar-Eclipse Pre _at_528 Eff Visible(),
Forward SS search
528
S0
A
Eclipse
S1
S2
11
Exogenous Conditions
Visible(A327)2300, 2320
Visible(A327)
2300
S0
A
Ex
S1
S2
12
Why is this Hard?
S0
A
Ex
S1
S2
Heuristic guidance!
13
Plan Graph Distance Heuristics
Construct plan graph Use DP to compute distance
estimates for goals
7
2
0
5
5
5
5
0
3
2
1
1
1
1
0
14
Adding Exogenous Events
7
2
0
X
5
X
5
5
5
0
0
X
1
1
1
1
0
Non-monotonicity
Actions become impossible Propositions go
away Distances get reset Mutex reappears
15
Reason Number 1

• Exogenous conditions events
• Representation
• Search guidance

16
Reason Number 2

• Exogenous conditions events
• Over-subscription problems
• Many observations/experiments
• Prioritized
• Limited time, energy, data storage, cryogen,

17
Rover Problem

• Given
• Set of goals g1, , gn with values v1, , vn
• Limited resources
• Objective
• Find a good plan

18
Rover Problem
5 rocks 5! orderings 25 rocks, can visit 5 6
million orderings
19
Which Goals?

• Backward search (Regression/POP/Graphplan)
• 2n goal sets
• Forward search
• guidance?

20
Plan Graph Resource Heuristics?
Construct plan graph Use DP to compute resource
estimates for goals
7
2
0
5
5
5
5
0
3
2
1
1
1
1
0
Choose goal set with greatest total utility given
the resources available
21
Plan Graph Resource Heuristics?
Construct plan graph Use DP to compute resource
estimates for goals
For Rover Pictures at all locations in 4
steps All estimates from current location
Problem Assumes independence between objectives
22
Plan Graph Utility Heuristics?
Construct plan graph Use DP to compute utility
estimates for actions
7
7
10
10
8
5
13
3
3
3
1
1
3
Choose actions that lead toward most utility
23
Propagating Utility Tables
5
5
5
25
15
e
q
g
A
B
(10, 15)
(10, 15)
p
t
1
C
r
2
(3, 3)
E
g
(2, 2)
D
s
t
(1, 5)
1
1
e
r
2
1
2
24
Plan Graph Utility Heuristics?
Construct plan graph Use DP to compute utility
estimates for actions
7
7
10
10
8
5
13
3
3
3
1
1
3
Problems Complex Assumes independence between
objectives
25
Reason Number 2

• Exogenous conditions events
• Over-subscription problems
• Poor search guidance

26
Reason Number 3

• Exogenous conditions events
• Over-subscription problems
• Uncertainty

Drive (-1)
Dig(5)
Visual servo (.2, -.15)
NIR
27
Reason Number 3

• Exogenous conditions events
• Over-subscription problems
• Uncertainty

?
10 ,1430
window
power
power
X
X
X
X
Drive (-1)
Dig(5)
Visual servo (.2, -.15)
NIR
28
Previous Work

Disjunction
Probability
CGP CMBP C-PLAN Fragplan
Non Observable
Problems
Buridan UDTPOP
Scalability STRIPS model of action no
concurrency no time no resources Discrete action
outcomes Utility
SENSp Cassandra PUCCINI SGP QBF-Plan GPT MBP
C-Buridan DTPOP C-MAXPLAN ZANDER Mahinur POMDPs
Partially Observable
JIC Plinth Weaver PGP MDPs
Fully Observable
WARPLAN-C CNLP
29
Can We Make it Discrete?
30
Can We Make it Discrete?
Picture
31
Can We Make it Discrete?
Collect
32
Optimal Value Function
E gt .1 Ah ? .05 Ah ? .02 Ah
E gt .6 Ah ? .2 Ah ? .2 Ah
E gt 10 Ah ? 5 Ah ? 2.5 Ah
E gt .02 Ah ? .01 Ah ? 0 Ah
t ? 900, 1430 ? 5s ? 1s
HiRes
V 10

t ? 1000, 1400 ? 600s ? 60s
E gt 3 Ah ? 2 Ah ? .5 Ah
? 1000s ? 500s
? 60s ? 1s
? 40s ? 20s
Drive (-2)
Dig(60)
Visual servo (.2, -.15)
NIR
Expected Value
V 100
t ? 900, 1600 ? 5s ? 1s
t ? 1000, 1350 ? 600s ? 60s
? 120s ? 20s
20
1320
15
Lo res
Rock finder
NIR
1340
V 50
V 5
10
1400
E gt .12 Ah ? .1 Ah ? .01 Ah
E gt .02 Ah ? .01 Ah ? 0 Ah
E gt 3 Ah ? 2 Ah ? .5 Ah
Power
5
1420
Start time
1440
33
Reason Number 3

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Continuous time ( resources)
• Continuous outcomes
• Utility

34
Reason Number 4

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Physical model
• Switches/valves

35
Opening Valves
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
36
Opening Valves
Pressure(T1)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
37
Opening Valves
Open(V1)
Pressure(T1)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
38
Opening Valves
Open(V1)
Pressure(T1)
Pressure(P2)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
39
Opening Valves
Pressure(P3) ??
Open(V1)
Pressure(T1)
Pressure(P2)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
40
Opening Valves
Pressure(P3)
Open(V1)
Open(V2)
Pressure(T1)
Pressure(P2)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
41
Opening Valves
Open(V2)
Pressure(T1)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
42
Opening Valves
Open(V1)
Open(V2)
Pressure(T1)
Pressure(P2)
Open (?valve, ?in, ?out) Pre Open(?valve) Eff
Open(?valve) Pressure(?in) ? Pressure(?out)
43
Reason Number 4

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Physical model
• Switches/valves
• Non-sequential activation

44
Reason Number 5

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan revision
• Rolling time horizon
• New goals
• Unexpected state

45
Replan from Scratch?

• Expensive
• Interferes with other unknown plans

46
Reason Number 5

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan revision
• Time constraints
• Minimize change

47
Reason Number 6

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan Revision
• External Reasoning

Drive (x, y)
max
Energy
min
48
Reason Number 6

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan Revision
• External Reasoning
• Resource models too complex

Simulator
Planner
Simulator
Drive (x, y)
max
Energy
min
49
Top 6 Reasons

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan Revision
• External Reasoning

50
Difficulty?

• Exogenous conditions events
• Over-subscription problems
• Uncertainty
• Ramifications
• Plan Revision
• External Reasoning

E M H E/M M ?
51
Questions?